II-NEW: An Experimental Platform for Investigating Energy-Performance Tradeoffs for Systems with Deep Memory Hierarchies

随着支持科学和工程的计算和数据基础设施的规模和复杂性的增长，电力成本正在成为成本，可靠性和整体可持续性方面的重要问题。因此，从新兴系统配置的应用角度理解功率/性能行为和权衡变得越来越重要，即，那些具有多核、深层内存层次结构和加速器的系统。该项目建立了一个仪器化的实验平台，支持这种理解，并使这一领域的研究和培训活动。具体而言，所提出的实验平台是由节点与深存储器架构，其中包含四个不同的水平：DRAM，基于PCIe的非易失性存储器，固态驱动器和旋转硬盘，除了加速器。该实验平台可以对大规模计算系统和数据中心及其支持的计算和数据密集型应用的功耗/性能行为进行实验探索，并为了解这些系统和应用的管理和优化提供独特的支持。它还支持多个领域的研究，包括：应用感知跨层管理、数据密集型科学工作流的功耗-性能权衡以及虚拟化云环境中深层内存层次结构的热影响。数据和计算密集型应用程序对广泛的领域越来越重要，开发大规模可持续平台和软件基础设施以支持这些应用程序的能力将对推动这些领域的研究和创新产生重大影响。开发的实验平台使关键的研究活动能够支持这一点。它提供了重要的见解，将影响当前和新兴数据和计算密集型应用所需的超大规模基础设施的实现和可持续性。该基础架构还为与电源管理、能效、数据管理、内存管理和虚拟化相关的不同领域的教育和培训提供了重要的基础架构。

As the scale and complexity of computing and data infrastructures supporting science and engineering grow, power costs are becoming important concerns in terms of costs, reliability and overall sustainability. As a result, it is becoming increasingly important to understand power/performance behaviors and tradeoffs from an application perspective for emerging system configuration, i.e., those with multiple cores, deep memory hierarchies and accelerators. This project builds an instrumented experimental platform that supports such an understanding, and enables research and training activities in this area. Specifically, the proposed experimental platform is composed of nodes with a deep memory architecture that contains four different levels: DRAM, PCIe-based non-volatile memory, solid-state drive and spinning hard disk, in addition to accelerators. Power metering is deployed as part of the infrastructure.The experimental platform enables the experimental exploration of the power/performance behaviors of large scale computing systems and datacenters as well as compute and data intensive application they support, and uniquely supports research toward understanding the management and optimization of these systems and applications. It also enables research in multiple areas, including: application-aware cross-layer management, power-performance tradeoffs for data-intensive scientific workflows and thermal implications of deep memory hierarchies in virtualized Cloud environments. Data and compute intensive applications are becoming increasingly critical to a wide range of domains, and the ability to develop large-scale and sustainable platforms and software infrastructure to support these applications will have significant impact in driving research and innovations in these domains. The developed experimental platform enables key research activities to support this. It provides important insights that will impact the realization and sustainability of very large-scale infrastructures necessary for current and emerging data and compute intensive applications. The infrastructure also provides an important infrastructure for education and training in different areas related to power management, energy efficiency, data management, memory management, and virtualization.